Agentoptosis: Cross-Scale Tradeoffs in Algorithmic Persistence

When a cell kills itself on purpose, or a bee stings and dies, or a worker ant walks out of the nest to die alone — these look like paradoxes of agency. How can an agent choose its own termination? This paper argues the paradox dissolves once you stop conflating four things that are genuinely distinct: what an agent is, what it models, what its objective serves, and what pattern is actually sustained by its actions.

The core move is separating the bearer of a policy (the thing that acts) from the target of that policy (the pattern whose persistence the objective serves). An agent's objective doesn't have to serve the agent itself. A cell can have rich self-monitoring machinery — tracking damage, infection, metabolic state — while running an objective calibrated to benefit the organism, not itself. Self-modeling and self-targeting are independent. This dissolves the apparent contradiction: the cell isn't being irrational; it's just other-targeting.

The paper introduces the term agentoptosis for the limiting case where a constituent both carries and executes its own termination machinery. It then builds a formal framework around four distinct patterns that must be tracked separately: the individual token $a_i$ (this cell, this worker), its recurring class $\mathcal{T}$ (the cell type, the worker role), the collective realization $M$ (this colony, this organism), and the recurring collective kind $\mathcal{L}(M)$ (the lineage, the species). A policy that destroys tokens can simultaneously preserve the class and benefit the collective — these are different observables that can have opposite signs. Snowflake yeast makes this concrete: apoptosis-like death promotes cluster fragmentation, which lowers persistence of the current cluster while raising persistence of the cluster lineage. Score it at $M$ and the sign is wrong; score it at $\mathcal{L}(M)$ and it's right.

Four engineering problems recur across all biological and artificial instances: which constituents are affordable to spend (those whose removal doesn't make the beneficiary harder to describe, because others already carry their contribution); what evidence should cross the termination threshold (a standard Bayes decision problem, with the threshold set by the ratio of false-positive to false-negative costs); why evasion doesn't take over (the stability problem — a death switch that cheaters can disable for free isn't stable, so relatedness, bottlenecks, or pleiotropic coupling must make evasion costly); and how the remains are handled (contained clearance, lytic signaling, terminal deployment, or uncontrolled dissolution — each selected by what the debris costs the beneficiary, not by the fact of death).

The paper is careful about what it doesn't claim. It explicitly declines to assert an adaptive account of human suicide, offering the framework instead as a vocabulary for stating competing hypotheses precisely — dysfunction in the agent's modeling or valuation machinery versus an evidential threshold responding to social signals like perceived burdensomeness. It also floats a conjecture that aging itself might be slow agentoptosis — a developmental schedule running the organism down to benefit a longer-lived lineage — but immediately subjects this to the same discipline it applies everywhere else: name the beneficiary, state the horizon, find the matched counterfactual, and explain why non-aging variants don't invade.